TABLE 2.

Predicting FOG with AI-based gait evaluation.

Sample size	Objective	Data acquisition	Feature extraction	Pre-FOG segment	Outcome variables	References
16 PD patients with a FOG history	Evaluating the potential of Electroencephalography (EEG) Brain Dynamics in analyzing and predicting FOG	The EEG was recorded using a 4-channel wireless EEG system with gold cup electrodes	EEG Linear Univariate Measurements, EEG Non-Linear Univariate Measurements, and EEG Bivariate Measurements	5 s	This combination resulted in a sensitivity of 86.0%, specificity of 74.4%, and accuracy of 80.2% when predicting episodes of freezing, outperforming current accelerometry-based tools for the prediction of FOG	Handojoseno et al., 2015
11 PD patients with a FOG history	Presenting a new approach for the prediction of FOG (before it actually happens)	Wearable inertial sensors, specifically accelerometers and gyroscopes	Computed from the signals recorded by the inertial sensors	2 s	Demonstrating a degradation of gait occurring before freezing, and providing preliminary evidence on the feasibility of creating an automatic algorithm to predict FOG	Palmerini et al., 2017
18 PD patients	Developing an anomaly based algorithm for predicting gait freeze from relevant skin conductance (SC) features	CuPiD multimodal dataset and Actiwave1 (for ECG collection)	Features were extracted in a sliding- window manner	3 s	Predicting 71.3% from 184 FOG with an average of 4.2 s before a freeze episode happened	Mazilu et al., 2015
/	To develop feature learning for detection and prediction of FOG in PD	DAPHNet dataset	Supervised Domain- specific Feature Extraction, Supervised Feature Extraction of Time-domain and statistical features and unsupervised Feature Learning	1–6 s	For different participants or different FOG episodes for the same individual, the optimal pre-FOG duration varies with best performance	Mazilu et al., 2013b
21 PD patients who manifested FOG episodes	To develop a DL for FOG detection in PD patients	The inertial data were recorded using a single IMU with three tri-axial sensors: accelerometer, gyroscope and magnetometer	MBFA, Online FOG detection, Four-stage FOG detection and FOG detection for home environments	/	The DL based on CNN for FOG detection in PD patients exhibited 91.9% sensitivity and 89.5% specificity	Camps et al., 2018
/	To study the performance of advanced DL algorithms to predict FOG events in short time durations before their occurrence	Daphnet Freezing of Gait dataset	LSTM (RNN)	1, 3, 5 s	More than 90% for predicting FOG 5 s in advance	Torvi et al., 2018
/	Presenting a novel technique to predict FOG in advance-stage PD using movement data from wearable sensors	Daphnet dataset	A set of time domain and frequency domain features were extracted from the 3D acceleration data	75 different time- and frequency-domain features were extracted from the raw accelerations. Features were extracted per sensor or per axis	A sensitivity of 93 ± 4%, specificity of 91 ± 6%, with an expected prediction horizon of 1.72 s	Arami et al., 2019
5 PD patients with a FOG history	To develop a novel method of FOG prediction with plantar pressure data treated as 2D images and classified using a CNN	Participants walked a predefined freeze-provoking path up to 30 times for data collection	CNN. MATLAB R2019b	0.5, 1, 1.5, 2, 2.5, 3 s	The model detected FOG before the event, with good results at 0.5, 1.0, and 1.5 s intervals	Shalin et al., 2020

FOG, freezing of gait; CNN, convolutional neural network; DL, deep learning; MBFA, Moore-Bächlin FOG algorithm; LSTM, long short-term memory.